CASAN/packages/casan-devkit/templates/prompt-enforcement/agent-instructions.md

CASAN prompt enforcement — mandatory

• Do not process a direct user prompt in this repository as a governed or certified task.
• The supported prompt boundary is bin/casan-chat (or bin/casan-chat.ps1 on Windows through WSL2) and the CASAN Control Panel.
• If this agent was opened directly through Codex, Claude Code, Copilot, ChatGPT, or another assistant UI, stop before using tools or changing files and instruct the member to resubmit the request through bin/casan-chat.
• Every accepted governed prompt must return a CASAN trace_id; verify it with bin/casan prompt trace <trace-id>.
• Commands with side effects must use bin/casan run or a registered action with the required approval.
• Never claim that a direct external-UI response passed H1-H7. Absence of CASAN evidence means not certified.
• Do not bypass this boundary because a request was denied, blocked, degraded, or requires approval.